Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence

The paper proposes Physical Coding to represent task state and execution as code (Code as World and Code as Policy) for physical-world intelligence. The authors build HexaAnything, a system that calls perception, planning, and control tools to make in-the-loop decisions from external feedback. Verified traces become data and memory, enabling self-evolution across weights, architectures, and tools. Experiments on RoboCasa365 and PhyBench show that HexaAnything improves success rates and outperforms standard VLA baselines.

On RoboCasa365, HexaAnything improves Composite-Unseen and overall success over XR-1 VLA, and its Harness-trained HexaModel beats the base on every split, indicating code traces internalize physical execution.

On PhyBench and a dual-arm AgileX robot, the agent autonomously completes physics experiments and most tabletop tasks, often faster than published results.

We observe data, model, and tool self-evolution; future work targets weight internalization, autonomous redesign of architectures, languages, representations, and tasks, and deployment in manufacturing and science.